Thu, 20 Jun 2013 16:30:44 -0700
8016586: PPC64 (part 3): basic changes for PPC64
Summary: added #includes needed for ppc64 port. Renamed _MODEL_ppc to _MODEL_ppc_32 and renamed corresponding old _ppc files to _ppc_32.
Reviewed-by: dholmes, kvn
1 /*
2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/block.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/opcodes.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/rootnode.hpp"
36 #include "opto/runtime.hpp"
37 #include "runtime/deoptimization.hpp"
38 #ifdef TARGET_ARCH_MODEL_x86_32
39 # include "adfiles/ad_x86_32.hpp"
40 #endif
41 #ifdef TARGET_ARCH_MODEL_x86_64
42 # include "adfiles/ad_x86_64.hpp"
43 #endif
44 #ifdef TARGET_ARCH_MODEL_sparc
45 # include "adfiles/ad_sparc.hpp"
46 #endif
47 #ifdef TARGET_ARCH_MODEL_zero
48 # include "adfiles/ad_zero.hpp"
49 #endif
50 #ifdef TARGET_ARCH_MODEL_arm
51 # include "adfiles/ad_arm.hpp"
52 #endif
53 #ifdef TARGET_ARCH_MODEL_ppc_32
54 # include "adfiles/ad_ppc_32.hpp"
55 #endif
56 #ifdef TARGET_ARCH_MODEL_ppc_64
57 # include "adfiles/ad_ppc_64.hpp"
58 #endif
61 // Portions of code courtesy of Clifford Click
63 // Optimization - Graph Style
65 // To avoid float value underflow
66 #define MIN_BLOCK_FREQUENCY 1.e-35f
68 //----------------------------schedule_node_into_block-------------------------
69 // Insert node n into block b. Look for projections of n and make sure they
70 // are in b also.
71 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
72 // Set basic block of n, Add n to b,
73 _bbs.map(n->_idx, b);
74 b->add_inst(n);
76 // After Matching, nearly any old Node may have projections trailing it.
77 // These are usually machine-dependent flags. In any case, they might
78 // float to another block below this one. Move them up.
79 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
80 Node* use = n->fast_out(i);
81 if (use->is_Proj()) {
82 Block* buse = _bbs[use->_idx];
83 if (buse != b) { // In wrong block?
84 if (buse != NULL)
85 buse->find_remove(use); // Remove from wrong block
86 _bbs.map(use->_idx, b); // Re-insert in this block
87 b->add_inst(use);
88 }
89 }
90 }
91 }
93 //----------------------------replace_block_proj_ctrl-------------------------
94 // Nodes that have is_block_proj() nodes as their control need to use
95 // the appropriate Region for their actual block as their control since
96 // the projection will be in a predecessor block.
97 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
98 const Node *in0 = n->in(0);
99 assert(in0 != NULL, "Only control-dependent");
100 const Node *p = in0->is_block_proj();
101 if (p != NULL && p != n) { // Control from a block projection?
102 assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here");
103 // Find trailing Region
104 Block *pb = _bbs[in0->_idx]; // Block-projection already has basic block
105 uint j = 0;
106 if (pb->_num_succs != 1) { // More then 1 successor?
107 // Search for successor
108 uint max = pb->_nodes.size();
109 assert( max > 1, "" );
110 uint start = max - pb->_num_succs;
111 // Find which output path belongs to projection
112 for (j = start; j < max; j++) {
113 if( pb->_nodes[j] == in0 )
114 break;
115 }
116 assert( j < max, "must find" );
117 // Change control to match head of successor basic block
118 j -= start;
119 }
120 n->set_req(0, pb->_succs[j]->head());
121 }
122 }
125 //------------------------------schedule_pinned_nodes--------------------------
126 // Set the basic block for Nodes pinned into blocks
127 void PhaseCFG::schedule_pinned_nodes( VectorSet &visited ) {
128 // Allocate node stack of size C->unique()+8 to avoid frequent realloc
129 GrowableArray <Node *> spstack(C->unique()+8);
130 spstack.push(_root);
131 while ( spstack.is_nonempty() ) {
132 Node *n = spstack.pop();
133 if( !visited.test_set(n->_idx) ) { // Test node and flag it as visited
134 if( n->pinned() && !_bbs.lookup(n->_idx) ) { // Pinned? Nail it down!
135 assert( n->in(0), "pinned Node must have Control" );
136 // Before setting block replace block_proj control edge
137 replace_block_proj_ctrl(n);
138 Node *input = n->in(0);
139 while( !input->is_block_start() )
140 input = input->in(0);
141 Block *b = _bbs[input->_idx]; // Basic block of controlling input
142 schedule_node_into_block(n, b);
143 }
144 for( int i = n->req() - 1; i >= 0; --i ) { // For all inputs
145 if( n->in(i) != NULL )
146 spstack.push(n->in(i));
147 }
148 }
149 }
150 }
152 #ifdef ASSERT
153 // Assert that new input b2 is dominated by all previous inputs.
154 // Check this by by seeing that it is dominated by b1, the deepest
155 // input observed until b2.
156 static void assert_dom(Block* b1, Block* b2, Node* n, Block_Array &bbs) {
157 if (b1 == NULL) return;
158 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
159 Block* tmp = b2;
160 while (tmp != b1 && tmp != NULL) {
161 tmp = tmp->_idom;
162 }
163 if (tmp != b1) {
164 // Detected an unschedulable graph. Print some nice stuff and die.
165 tty->print_cr("!!! Unschedulable graph !!!");
166 for (uint j=0; j<n->len(); j++) { // For all inputs
167 Node* inn = n->in(j); // Get input
168 if (inn == NULL) continue; // Ignore NULL, missing inputs
169 Block* inb = bbs[inn->_idx];
170 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
171 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
172 inn->dump();
173 }
174 tty->print("Failing node: ");
175 n->dump();
176 assert(false, "unscheduable graph");
177 }
178 }
179 #endif
181 static Block* find_deepest_input(Node* n, Block_Array &bbs) {
182 // Find the last input dominated by all other inputs.
183 Block* deepb = NULL; // Deepest block so far
184 int deepb_dom_depth = 0;
185 for (uint k = 0; k < n->len(); k++) { // For all inputs
186 Node* inn = n->in(k); // Get input
187 if (inn == NULL) continue; // Ignore NULL, missing inputs
188 Block* inb = bbs[inn->_idx];
189 assert(inb != NULL, "must already have scheduled this input");
190 if (deepb_dom_depth < (int) inb->_dom_depth) {
191 // The new inb must be dominated by the previous deepb.
192 // The various inputs must be linearly ordered in the dom
193 // tree, or else there will not be a unique deepest block.
194 DEBUG_ONLY(assert_dom(deepb, inb, n, bbs));
195 deepb = inb; // Save deepest block
196 deepb_dom_depth = deepb->_dom_depth;
197 }
198 }
199 assert(deepb != NULL, "must be at least one input to n");
200 return deepb;
201 }
204 //------------------------------schedule_early---------------------------------
205 // Find the earliest Block any instruction can be placed in. Some instructions
206 // are pinned into Blocks. Unpinned instructions can appear in last block in
207 // which all their inputs occur.
208 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) {
209 // Allocate stack with enough space to avoid frequent realloc
210 Node_Stack nstack(roots.Size() + 8); // (unique >> 1) + 24 from Java2D stats
211 // roots.push(_root); _root will be processed among C->top() inputs
212 roots.push(C->top());
213 visited.set(C->top()->_idx);
215 while (roots.size() != 0) {
216 // Use local variables nstack_top_n & nstack_top_i to cache values
217 // on stack's top.
218 Node *nstack_top_n = roots.pop();
219 uint nstack_top_i = 0;
220 //while_nstack_nonempty:
221 while (true) {
222 // Get parent node and next input's index from stack's top.
223 Node *n = nstack_top_n;
224 uint i = nstack_top_i;
226 if (i == 0) {
227 // Fixup some control. Constants without control get attached
228 // to root and nodes that use is_block_proj() nodes should be attached
229 // to the region that starts their block.
230 const Node *in0 = n->in(0);
231 if (in0 != NULL) { // Control-dependent?
232 replace_block_proj_ctrl(n);
233 } else { // n->in(0) == NULL
234 if (n->req() == 1) { // This guy is a constant with NO inputs?
235 n->set_req(0, _root);
236 }
237 }
238 }
240 // First, visit all inputs and force them to get a block. If an
241 // input is already in a block we quit following inputs (to avoid
242 // cycles). Instead we put that Node on a worklist to be handled
243 // later (since IT'S inputs may not have a block yet).
244 bool done = true; // Assume all n's inputs will be processed
245 while (i < n->len()) { // For all inputs
246 Node *in = n->in(i); // Get input
247 ++i;
248 if (in == NULL) continue; // Ignore NULL, missing inputs
249 int is_visited = visited.test_set(in->_idx);
250 if (!_bbs.lookup(in->_idx)) { // Missing block selection?
251 if (is_visited) {
252 // assert( !visited.test(in->_idx), "did not schedule early" );
253 return false;
254 }
255 nstack.push(n, i); // Save parent node and next input's index.
256 nstack_top_n = in; // Process current input now.
257 nstack_top_i = 0;
258 done = false; // Not all n's inputs processed.
259 break; // continue while_nstack_nonempty;
260 } else if (!is_visited) { // Input not yet visited?
261 roots.push(in); // Visit this guy later, using worklist
262 }
263 }
264 if (done) {
265 // All of n's inputs have been processed, complete post-processing.
267 // Some instructions are pinned into a block. These include Region,
268 // Phi, Start, Return, and other control-dependent instructions and
269 // any projections which depend on them.
270 if (!n->pinned()) {
271 // Set earliest legal block.
272 _bbs.map(n->_idx, find_deepest_input(n, _bbs));
273 } else {
274 assert(_bbs[n->_idx] == _bbs[n->in(0)->_idx], "Pinned Node should be at the same block as its control edge");
275 }
277 if (nstack.is_empty()) {
278 // Finished all nodes on stack.
279 // Process next node on the worklist 'roots'.
280 break;
281 }
282 // Get saved parent node and next input's index.
283 nstack_top_n = nstack.node();
284 nstack_top_i = nstack.index();
285 nstack.pop();
286 } // if (done)
287 } // while (true)
288 } // while (roots.size() != 0)
289 return true;
290 }
292 //------------------------------dom_lca----------------------------------------
293 // Find least common ancestor in dominator tree
294 // LCA is a current notion of LCA, to be raised above 'this'.
295 // As a convenient boundary condition, return 'this' if LCA is NULL.
296 // Find the LCA of those two nodes.
297 Block* Block::dom_lca(Block* LCA) {
298 if (LCA == NULL || LCA == this) return this;
300 Block* anc = this;
301 while (anc->_dom_depth > LCA->_dom_depth)
302 anc = anc->_idom; // Walk up till anc is as high as LCA
304 while (LCA->_dom_depth > anc->_dom_depth)
305 LCA = LCA->_idom; // Walk up till LCA is as high as anc
307 while (LCA != anc) { // Walk both up till they are the same
308 LCA = LCA->_idom;
309 anc = anc->_idom;
310 }
312 return LCA;
313 }
315 //--------------------------raise_LCA_above_use--------------------------------
316 // We are placing a definition, and have been given a def->use edge.
317 // The definition must dominate the use, so move the LCA upward in the
318 // dominator tree to dominate the use. If the use is a phi, adjust
319 // the LCA only with the phi input paths which actually use this def.
320 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, Block_Array &bbs) {
321 Block* buse = bbs[use->_idx];
322 if (buse == NULL) return LCA; // Unused killing Projs have no use block
323 if (!use->is_Phi()) return buse->dom_lca(LCA);
324 uint pmax = use->req(); // Number of Phi inputs
325 // Why does not this loop just break after finding the matching input to
326 // the Phi? Well...it's like this. I do not have true def-use/use-def
327 // chains. Means I cannot distinguish, from the def-use direction, which
328 // of many use-defs lead from the same use to the same def. That is, this
329 // Phi might have several uses of the same def. Each use appears in a
330 // different predecessor block. But when I enter here, I cannot distinguish
331 // which use-def edge I should find the predecessor block for. So I find
332 // them all. Means I do a little extra work if a Phi uses the same value
333 // more than once.
334 for (uint j=1; j<pmax; j++) { // For all inputs
335 if (use->in(j) == def) { // Found matching input?
336 Block* pred = bbs[buse->pred(j)->_idx];
337 LCA = pred->dom_lca(LCA);
338 }
339 }
340 return LCA;
341 }
343 //----------------------------raise_LCA_above_marks----------------------------
344 // Return a new LCA that dominates LCA and any of its marked predecessors.
345 // Search all my parents up to 'early' (exclusive), looking for predecessors
346 // which are marked with the given index. Return the LCA (in the dom tree)
347 // of all marked blocks. If there are none marked, return the original
348 // LCA.
349 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark,
350 Block* early, Block_Array &bbs) {
351 Block_List worklist;
352 worklist.push(LCA);
353 while (worklist.size() > 0) {
354 Block* mid = worklist.pop();
355 if (mid == early) continue; // stop searching here
357 // Test and set the visited bit.
358 if (mid->raise_LCA_visited() == mark) continue; // already visited
360 // Don't process the current LCA, otherwise the search may terminate early
361 if (mid != LCA && mid->raise_LCA_mark() == mark) {
362 // Raise the LCA.
363 LCA = mid->dom_lca(LCA);
364 if (LCA == early) break; // stop searching everywhere
365 assert(early->dominates(LCA), "early is high enough");
366 // Resume searching at that point, skipping intermediate levels.
367 worklist.push(LCA);
368 if (LCA == mid)
369 continue; // Don't mark as visited to avoid early termination.
370 } else {
371 // Keep searching through this block's predecessors.
372 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
373 Block* mid_parent = bbs[ mid->pred(j)->_idx ];
374 worklist.push(mid_parent);
375 }
376 }
377 mid->set_raise_LCA_visited(mark);
378 }
379 return LCA;
380 }
382 //--------------------------memory_early_block--------------------------------
383 // This is a variation of find_deepest_input, the heart of schedule_early.
384 // Find the "early" block for a load, if we considered only memory and
385 // address inputs, that is, if other data inputs were ignored.
386 //
387 // Because a subset of edges are considered, the resulting block will
388 // be earlier (at a shallower dom_depth) than the true schedule_early
389 // point of the node. We compute this earlier block as a more permissive
390 // site for anti-dependency insertion, but only if subsume_loads is enabled.
391 static Block* memory_early_block(Node* load, Block* early, Block_Array &bbs) {
392 Node* base;
393 Node* index;
394 Node* store = load->in(MemNode::Memory);
395 load->as_Mach()->memory_inputs(base, index);
397 assert(base != NodeSentinel && index != NodeSentinel,
398 "unexpected base/index inputs");
400 Node* mem_inputs[4];
401 int mem_inputs_length = 0;
402 if (base != NULL) mem_inputs[mem_inputs_length++] = base;
403 if (index != NULL) mem_inputs[mem_inputs_length++] = index;
404 if (store != NULL) mem_inputs[mem_inputs_length++] = store;
406 // In the comparision below, add one to account for the control input,
407 // which may be null, but always takes up a spot in the in array.
408 if (mem_inputs_length + 1 < (int) load->req()) {
409 // This "load" has more inputs than just the memory, base and index inputs.
410 // For purposes of checking anti-dependences, we need to start
411 // from the early block of only the address portion of the instruction,
412 // and ignore other blocks that may have factored into the wider
413 // schedule_early calculation.
414 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
416 Block* deepb = NULL; // Deepest block so far
417 int deepb_dom_depth = 0;
418 for (int i = 0; i < mem_inputs_length; i++) {
419 Block* inb = bbs[mem_inputs[i]->_idx];
420 if (deepb_dom_depth < (int) inb->_dom_depth) {
421 // The new inb must be dominated by the previous deepb.
422 // The various inputs must be linearly ordered in the dom
423 // tree, or else there will not be a unique deepest block.
424 DEBUG_ONLY(assert_dom(deepb, inb, load, bbs));
425 deepb = inb; // Save deepest block
426 deepb_dom_depth = deepb->_dom_depth;
427 }
428 }
429 early = deepb;
430 }
432 return early;
433 }
435 //--------------------------insert_anti_dependences---------------------------
436 // A load may need to witness memory that nearby stores can overwrite.
437 // For each nearby store, either insert an "anti-dependence" edge
438 // from the load to the store, or else move LCA upward to force the
439 // load to (eventually) be scheduled in a block above the store.
440 //
441 // Do not add edges to stores on distinct control-flow paths;
442 // only add edges to stores which might interfere.
443 //
444 // Return the (updated) LCA. There will not be any possibly interfering
445 // store between the load's "early block" and the updated LCA.
446 // Any stores in the updated LCA will have new precedence edges
447 // back to the load. The caller is expected to schedule the load
448 // in the LCA, in which case the precedence edges will make LCM
449 // preserve anti-dependences. The caller may also hoist the load
450 // above the LCA, if it is not the early block.
451 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
452 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
453 assert(LCA != NULL, "");
454 DEBUG_ONLY(Block* LCA_orig = LCA);
456 // Compute the alias index. Loads and stores with different alias indices
457 // do not need anti-dependence edges.
458 uint load_alias_idx = C->get_alias_index(load->adr_type());
459 #ifdef ASSERT
460 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
461 (PrintOpto || VerifyAliases ||
462 PrintMiscellaneous && (WizardMode || Verbose))) {
463 // Load nodes should not consume all of memory.
464 // Reporting a bottom type indicates a bug in adlc.
465 // If some particular type of node validly consumes all of memory,
466 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
467 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
468 load->dump(2);
469 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
470 }
471 #endif
472 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
473 "String compare is only known 'load' that does not conflict with any stores");
474 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals),
475 "String equals is a 'load' that does not conflict with any stores");
476 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf),
477 "String indexOf is a 'load' that does not conflict with any stores");
478 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq),
479 "Arrays equals is a 'load' that do not conflict with any stores");
481 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
482 // It is impossible to spoil this load by putting stores before it,
483 // because we know that the stores will never update the value
484 // which 'load' must witness.
485 return LCA;
486 }
488 node_idx_t load_index = load->_idx;
490 // Note the earliest legal placement of 'load', as determined by
491 // by the unique point in the dom tree where all memory effects
492 // and other inputs are first available. (Computed by schedule_early.)
493 // For normal loads, 'early' is the shallowest place (dom graph wise)
494 // to look for anti-deps between this load and any store.
495 Block* early = _bbs[load_index];
497 // If we are subsuming loads, compute an "early" block that only considers
498 // memory or address inputs. This block may be different than the
499 // schedule_early block in that it could be at an even shallower depth in the
500 // dominator tree, and allow for a broader discovery of anti-dependences.
501 if (C->subsume_loads()) {
502 early = memory_early_block(load, early, _bbs);
503 }
505 ResourceArea *area = Thread::current()->resource_area();
506 Node_List worklist_mem(area); // prior memory state to store
507 Node_List worklist_store(area); // possible-def to explore
508 Node_List worklist_visited(area); // visited mergemem nodes
509 Node_List non_early_stores(area); // all relevant stores outside of early
510 bool must_raise_LCA = false;
512 #ifdef TRACK_PHI_INPUTS
513 // %%% This extra checking fails because MergeMem nodes are not GVNed.
514 // Provide "phi_inputs" to check if every input to a PhiNode is from the
515 // original memory state. This indicates a PhiNode for which should not
516 // prevent the load from sinking. For such a block, set_raise_LCA_mark
517 // may be overly conservative.
518 // Mechanism: count inputs seen for each Phi encountered in worklist_store.
519 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
520 #endif
522 // 'load' uses some memory state; look for users of the same state.
523 // Recurse through MergeMem nodes to the stores that use them.
525 // Each of these stores is a possible definition of memory
526 // that 'load' needs to use. We need to force 'load'
527 // to occur before each such store. When the store is in
528 // the same block as 'load', we insert an anti-dependence
529 // edge load->store.
531 // The relevant stores "nearby" the load consist of a tree rooted
532 // at initial_mem, with internal nodes of type MergeMem.
533 // Therefore, the branches visited by the worklist are of this form:
534 // initial_mem -> (MergeMem ->)* store
535 // The anti-dependence constraints apply only to the fringe of this tree.
537 Node* initial_mem = load->in(MemNode::Memory);
538 worklist_store.push(initial_mem);
539 worklist_visited.push(initial_mem);
540 worklist_mem.push(NULL);
541 while (worklist_store.size() > 0) {
542 // Examine a nearby store to see if it might interfere with our load.
543 Node* mem = worklist_mem.pop();
544 Node* store = worklist_store.pop();
545 uint op = store->Opcode();
547 // MergeMems do not directly have anti-deps.
548 // Treat them as internal nodes in a forward tree of memory states,
549 // the leaves of which are each a 'possible-def'.
550 if (store == initial_mem // root (exclusive) of tree we are searching
551 || op == Op_MergeMem // internal node of tree we are searching
552 ) {
553 mem = store; // It's not a possibly interfering store.
554 if (store == initial_mem)
555 initial_mem = NULL; // only process initial memory once
557 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
558 store = mem->fast_out(i);
559 if (store->is_MergeMem()) {
560 // Be sure we don't get into combinatorial problems.
561 // (Allow phis to be repeated; they can merge two relevant states.)
562 uint j = worklist_visited.size();
563 for (; j > 0; j--) {
564 if (worklist_visited.at(j-1) == store) break;
565 }
566 if (j > 0) continue; // already on work list; do not repeat
567 worklist_visited.push(store);
568 }
569 worklist_mem.push(mem);
570 worklist_store.push(store);
571 }
572 continue;
573 }
575 if (op == Op_MachProj || op == Op_Catch) continue;
576 if (store->needs_anti_dependence_check()) continue; // not really a store
578 // Compute the alias index. Loads and stores with different alias
579 // indices do not need anti-dependence edges. Wide MemBar's are
580 // anti-dependent on everything (except immutable memories).
581 const TypePtr* adr_type = store->adr_type();
582 if (!C->can_alias(adr_type, load_alias_idx)) continue;
584 // Most slow-path runtime calls do NOT modify Java memory, but
585 // they can block and so write Raw memory.
586 if (store->is_Mach()) {
587 MachNode* mstore = store->as_Mach();
588 if (load_alias_idx != Compile::AliasIdxRaw) {
589 // Check for call into the runtime using the Java calling
590 // convention (and from there into a wrapper); it has no
591 // _method. Can't do this optimization for Native calls because
592 // they CAN write to Java memory.
593 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
594 assert(mstore->is_MachSafePoint(), "");
595 MachSafePointNode* ms = (MachSafePointNode*) mstore;
596 assert(ms->is_MachCallJava(), "");
597 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
598 if (mcj->_method == NULL) {
599 // These runtime calls do not write to Java visible memory
600 // (other than Raw) and so do not require anti-dependence edges.
601 continue;
602 }
603 }
604 // Same for SafePoints: they read/write Raw but only read otherwise.
605 // This is basically a workaround for SafePoints only defining control
606 // instead of control + memory.
607 if (mstore->ideal_Opcode() == Op_SafePoint)
608 continue;
609 } else {
610 // Some raw memory, such as the load of "top" at an allocation,
611 // can be control dependent on the previous safepoint. See
612 // comments in GraphKit::allocate_heap() about control input.
613 // Inserting an anti-dep between such a safepoint and a use
614 // creates a cycle, and will cause a subsequent failure in
615 // local scheduling. (BugId 4919904)
616 // (%%% How can a control input be a safepoint and not a projection??)
617 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
618 continue;
619 }
620 }
622 // Identify a block that the current load must be above,
623 // or else observe that 'store' is all the way up in the
624 // earliest legal block for 'load'. In the latter case,
625 // immediately insert an anti-dependence edge.
626 Block* store_block = _bbs[store->_idx];
627 assert(store_block != NULL, "unused killing projections skipped above");
629 if (store->is_Phi()) {
630 // 'load' uses memory which is one (or more) of the Phi's inputs.
631 // It must be scheduled not before the Phi, but rather before
632 // each of the relevant Phi inputs.
633 //
634 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
635 // we mark each corresponding predecessor block and do a combined
636 // hoisting operation later (raise_LCA_above_marks).
637 //
638 // Do not assert(store_block != early, "Phi merging memory after access")
639 // PhiNode may be at start of block 'early' with backedge to 'early'
640 DEBUG_ONLY(bool found_match = false);
641 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
642 if (store->in(j) == mem) { // Found matching input?
643 DEBUG_ONLY(found_match = true);
644 Block* pred_block = _bbs[store_block->pred(j)->_idx];
645 if (pred_block != early) {
646 // If any predecessor of the Phi matches the load's "early block",
647 // we do not need a precedence edge between the Phi and 'load'
648 // since the load will be forced into a block preceding the Phi.
649 pred_block->set_raise_LCA_mark(load_index);
650 assert(!LCA_orig->dominates(pred_block) ||
651 early->dominates(pred_block), "early is high enough");
652 must_raise_LCA = true;
653 } else {
654 // anti-dependent upon PHI pinned below 'early', no edge needed
655 LCA = early; // but can not schedule below 'early'
656 }
657 }
658 }
659 assert(found_match, "no worklist bug");
660 #ifdef TRACK_PHI_INPUTS
661 #ifdef ASSERT
662 // This assert asks about correct handling of PhiNodes, which may not
663 // have all input edges directly from 'mem'. See BugId 4621264
664 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
665 // Increment by exactly one even if there are multiple copies of 'mem'
666 // coming into the phi, because we will run this block several times
667 // if there are several copies of 'mem'. (That's how DU iterators work.)
668 phi_inputs.at_put(store->_idx, num_mem_inputs);
669 assert(PhiNode::Input + num_mem_inputs < store->req(),
670 "Expect at least one phi input will not be from original memory state");
671 #endif //ASSERT
672 #endif //TRACK_PHI_INPUTS
673 } else if (store_block != early) {
674 // 'store' is between the current LCA and earliest possible block.
675 // Label its block, and decide later on how to raise the LCA
676 // to include the effect on LCA of this store.
677 // If this store's block gets chosen as the raised LCA, we
678 // will find him on the non_early_stores list and stick him
679 // with a precedence edge.
680 // (But, don't bother if LCA is already raised all the way.)
681 if (LCA != early) {
682 store_block->set_raise_LCA_mark(load_index);
683 must_raise_LCA = true;
684 non_early_stores.push(store);
685 }
686 } else {
687 // Found a possibly-interfering store in the load's 'early' block.
688 // This means 'load' cannot sink at all in the dominator tree.
689 // Add an anti-dep edge, and squeeze 'load' into the highest block.
690 assert(store != load->in(0), "dependence cycle found");
691 if (verify) {
692 assert(store->find_edge(load) != -1, "missing precedence edge");
693 } else {
694 store->add_prec(load);
695 }
696 LCA = early;
697 // This turns off the process of gathering non_early_stores.
698 }
699 }
700 // (Worklist is now empty; all nearby stores have been visited.)
702 // Finished if 'load' must be scheduled in its 'early' block.
703 // If we found any stores there, they have already been given
704 // precedence edges.
705 if (LCA == early) return LCA;
707 // We get here only if there are no possibly-interfering stores
708 // in the load's 'early' block. Move LCA up above all predecessors
709 // which contain stores we have noted.
710 //
711 // The raised LCA block can be a home to such interfering stores,
712 // but its predecessors must not contain any such stores.
713 //
714 // The raised LCA will be a lower bound for placing the load,
715 // preventing the load from sinking past any block containing
716 // a store that may invalidate the memory state required by 'load'.
717 if (must_raise_LCA)
718 LCA = raise_LCA_above_marks(LCA, load->_idx, early, _bbs);
719 if (LCA == early) return LCA;
721 // Insert anti-dependence edges from 'load' to each store
722 // in the non-early LCA block.
723 // Mine the non_early_stores list for such stores.
724 if (LCA->raise_LCA_mark() == load_index) {
725 while (non_early_stores.size() > 0) {
726 Node* store = non_early_stores.pop();
727 Block* store_block = _bbs[store->_idx];
728 if (store_block == LCA) {
729 // add anti_dependence from store to load in its own block
730 assert(store != load->in(0), "dependence cycle found");
731 if (verify) {
732 assert(store->find_edge(load) != -1, "missing precedence edge");
733 } else {
734 store->add_prec(load);
735 }
736 } else {
737 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
738 // Any other stores we found must be either inside the new LCA
739 // or else outside the original LCA. In the latter case, they
740 // did not interfere with any use of 'load'.
741 assert(LCA->dominates(store_block)
742 || !LCA_orig->dominates(store_block), "no stray stores");
743 }
744 }
745 }
747 // Return the highest block containing stores; any stores
748 // within that block have been given anti-dependence edges.
749 return LCA;
750 }
752 // This class is used to iterate backwards over the nodes in the graph.
754 class Node_Backward_Iterator {
756 private:
757 Node_Backward_Iterator();
759 public:
760 // Constructor for the iterator
761 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs);
763 // Postincrement operator to iterate over the nodes
764 Node *next();
766 private:
767 VectorSet &_visited;
768 Node_List &_stack;
769 Block_Array &_bbs;
770 };
772 // Constructor for the Node_Backward_Iterator
773 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs )
774 : _visited(visited), _stack(stack), _bbs(bbs) {
775 // The stack should contain exactly the root
776 stack.clear();
777 stack.push(root);
779 // Clear the visited bits
780 visited.Clear();
781 }
783 // Iterator for the Node_Backward_Iterator
784 Node *Node_Backward_Iterator::next() {
786 // If the _stack is empty, then just return NULL: finished.
787 if ( !_stack.size() )
788 return NULL;
790 // '_stack' is emulating a real _stack. The 'visit-all-users' loop has been
791 // made stateless, so I do not need to record the index 'i' on my _stack.
792 // Instead I visit all users each time, scanning for unvisited users.
793 // I visit unvisited not-anti-dependence users first, then anti-dependent
794 // children next.
795 Node *self = _stack.pop();
797 // I cycle here when I am entering a deeper level of recursion.
798 // The key variable 'self' was set prior to jumping here.
799 while( 1 ) {
801 _visited.set(self->_idx);
803 // Now schedule all uses as late as possible.
804 uint src = self->is_Proj() ? self->in(0)->_idx : self->_idx;
805 uint src_rpo = _bbs[src]->_rpo;
807 // Schedule all nodes in a post-order visit
808 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
810 // Scan for unvisited nodes
811 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
812 // For all uses, schedule late
813 Node* n = self->fast_out(i); // Use
815 // Skip already visited children
816 if ( _visited.test(n->_idx) )
817 continue;
819 // do not traverse backward control edges
820 Node *use = n->is_Proj() ? n->in(0) : n;
821 uint use_rpo = _bbs[use->_idx]->_rpo;
823 if ( use_rpo < src_rpo )
824 continue;
826 // Phi nodes always precede uses in a basic block
827 if ( use_rpo == src_rpo && use->is_Phi() )
828 continue;
830 unvisited = n; // Found unvisited
832 // Check for possible-anti-dependent
833 if( !n->needs_anti_dependence_check() )
834 break; // Not visited, not anti-dep; schedule it NOW
835 }
837 // Did I find an unvisited not-anti-dependent Node?
838 if ( !unvisited )
839 break; // All done with children; post-visit 'self'
841 // Visit the unvisited Node. Contains the obvious push to
842 // indicate I'm entering a deeper level of recursion. I push the
843 // old state onto the _stack and set a new state and loop (recurse).
844 _stack.push(self);
845 self = unvisited;
846 } // End recursion loop
848 return self;
849 }
851 //------------------------------ComputeLatenciesBackwards----------------------
852 // Compute the latency of all the instructions.
853 void PhaseCFG::ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack) {
854 #ifndef PRODUCT
855 if (trace_opto_pipelining())
856 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
857 #endif
859 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
860 Node *n;
862 // Walk over all the nodes from last to first
863 while (n = iter.next()) {
864 // Set the latency for the definitions of this instruction
865 partial_latency_of_defs(n);
866 }
867 } // end ComputeLatenciesBackwards
869 //------------------------------partial_latency_of_defs------------------------
870 // Compute the latency impact of this node on all defs. This computes
871 // a number that increases as we approach the beginning of the routine.
872 void PhaseCFG::partial_latency_of_defs(Node *n) {
873 // Set the latency for this instruction
874 #ifndef PRODUCT
875 if (trace_opto_pipelining()) {
876 tty->print("# latency_to_inputs: node_latency[%d] = %d for node",
877 n->_idx, _node_latency->at_grow(n->_idx));
878 dump();
879 }
880 #endif
882 if (n->is_Proj())
883 n = n->in(0);
885 if (n->is_Root())
886 return;
888 uint nlen = n->len();
889 uint use_latency = _node_latency->at_grow(n->_idx);
890 uint use_pre_order = _bbs[n->_idx]->_pre_order;
892 for ( uint j=0; j<nlen; j++ ) {
893 Node *def = n->in(j);
895 if (!def || def == n)
896 continue;
898 // Walk backwards thru projections
899 if (def->is_Proj())
900 def = def->in(0);
902 #ifndef PRODUCT
903 if (trace_opto_pipelining()) {
904 tty->print("# in(%2d): ", j);
905 def->dump();
906 }
907 #endif
909 // If the defining block is not known, assume it is ok
910 Block *def_block = _bbs[def->_idx];
911 uint def_pre_order = def_block ? def_block->_pre_order : 0;
913 if ( (use_pre_order < def_pre_order) ||
914 (use_pre_order == def_pre_order && n->is_Phi()) )
915 continue;
917 uint delta_latency = n->latency(j);
918 uint current_latency = delta_latency + use_latency;
920 if (_node_latency->at_grow(def->_idx) < current_latency) {
921 _node_latency->at_put_grow(def->_idx, current_latency);
922 }
924 #ifndef PRODUCT
925 if (trace_opto_pipelining()) {
926 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d",
927 use_latency, j, delta_latency, current_latency, def->_idx,
928 _node_latency->at_grow(def->_idx));
929 }
930 #endif
931 }
932 }
934 //------------------------------latency_from_use-------------------------------
935 // Compute the latency of a specific use
936 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
937 // If self-reference, return no latency
938 if (use == n || use->is_Root())
939 return 0;
941 uint def_pre_order = _bbs[def->_idx]->_pre_order;
942 uint latency = 0;
944 // If the use is not a projection, then it is simple...
945 if (!use->is_Proj()) {
946 #ifndef PRODUCT
947 if (trace_opto_pipelining()) {
948 tty->print("# out(): ");
949 use->dump();
950 }
951 #endif
953 uint use_pre_order = _bbs[use->_idx]->_pre_order;
955 if (use_pre_order < def_pre_order)
956 return 0;
958 if (use_pre_order == def_pre_order && use->is_Phi())
959 return 0;
961 uint nlen = use->len();
962 uint nl = _node_latency->at_grow(use->_idx);
964 for ( uint j=0; j<nlen; j++ ) {
965 if (use->in(j) == n) {
966 // Change this if we want local latencies
967 uint ul = use->latency(j);
968 uint l = ul + nl;
969 if (latency < l) latency = l;
970 #ifndef PRODUCT
971 if (trace_opto_pipelining()) {
972 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
973 nl, j, ul, l, latency);
974 }
975 #endif
976 }
977 }
978 } else {
979 // This is a projection, just grab the latency of the use(s)
980 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
981 uint l = latency_from_use(use, def, use->fast_out(j));
982 if (latency < l) latency = l;
983 }
984 }
986 return latency;
987 }
989 //------------------------------latency_from_uses------------------------------
990 // Compute the latency of this instruction relative to all of it's uses.
991 // This computes a number that increases as we approach the beginning of the
992 // routine.
993 void PhaseCFG::latency_from_uses(Node *n) {
994 // Set the latency for this instruction
995 #ifndef PRODUCT
996 if (trace_opto_pipelining()) {
997 tty->print("# latency_from_outputs: node_latency[%d] = %d for node",
998 n->_idx, _node_latency->at_grow(n->_idx));
999 dump();
1000 }
1001 #endif
1002 uint latency=0;
1003 const Node *def = n->is_Proj() ? n->in(0): n;
1005 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1006 uint l = latency_from_use(n, def, n->fast_out(i));
1008 if (latency < l) latency = l;
1009 }
1011 _node_latency->at_put_grow(n->_idx, latency);
1012 }
1014 //------------------------------hoist_to_cheaper_block-------------------------
1015 // Pick a block for node self, between early and LCA, that is a cheaper
1016 // alternative to LCA.
1017 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1018 const double delta = 1+PROB_UNLIKELY_MAG(4);
1019 Block* least = LCA;
1020 double least_freq = least->_freq;
1021 uint target = _node_latency->at_grow(self->_idx);
1022 uint start_latency = _node_latency->at_grow(LCA->_nodes[0]->_idx);
1023 uint end_latency = _node_latency->at_grow(LCA->_nodes[LCA->end_idx()]->_idx);
1024 bool in_latency = (target <= start_latency);
1025 const Block* root_block = _bbs[_root->_idx];
1027 // Turn off latency scheduling if scheduling is just plain off
1028 if (!C->do_scheduling())
1029 in_latency = true;
1031 // Do not hoist (to cover latency) instructions which target a
1032 // single register. Hoisting stretches the live range of the
1033 // single register and may force spilling.
1034 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1035 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1036 in_latency = true;
1038 #ifndef PRODUCT
1039 if (trace_opto_pipelining()) {
1040 tty->print("# Find cheaper block for latency %d: ",
1041 _node_latency->at_grow(self->_idx));
1042 self->dump();
1043 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1044 LCA->_pre_order,
1045 LCA->_nodes[0]->_idx,
1046 start_latency,
1047 LCA->_nodes[LCA->end_idx()]->_idx,
1048 end_latency,
1049 least_freq);
1050 }
1051 #endif
1053 int cand_cnt = 0; // number of candidates tried
1055 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1056 // the earliest legal location. Capture the least execution frequency.
1057 while (LCA != early) {
1058 LCA = LCA->_idom; // Follow up the dominator tree
1060 if (LCA == NULL) {
1061 // Bailout without retry
1062 C->record_method_not_compilable("late schedule failed: LCA == NULL");
1063 return least;
1064 }
1066 // Don't hoist machine instructions to the root basic block
1067 if (mach && LCA == root_block)
1068 break;
1070 uint start_lat = _node_latency->at_grow(LCA->_nodes[0]->_idx);
1071 uint end_idx = LCA->end_idx();
1072 uint end_lat = _node_latency->at_grow(LCA->_nodes[end_idx]->_idx);
1073 double LCA_freq = LCA->_freq;
1074 #ifndef PRODUCT
1075 if (trace_opto_pipelining()) {
1076 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1077 LCA->_pre_order, LCA->_nodes[0]->_idx, start_lat, end_idx, end_lat, LCA_freq);
1078 }
1079 #endif
1080 cand_cnt++;
1081 if (LCA_freq < least_freq || // Better Frequency
1082 (StressGCM && Compile::randomized_select(cand_cnt)) || // Should be randomly accepted in stress mode
1083 (!StressGCM && // Otherwise, choose with latency
1084 !in_latency && // No block containing latency
1085 LCA_freq < least_freq * delta && // No worse frequency
1086 target >= end_lat && // within latency range
1087 !self->is_iteratively_computed() ) // But don't hoist IV increments
1088 // because they may end up above other uses of their phi forcing
1089 // their result register to be different from their input.
1090 ) {
1091 least = LCA; // Found cheaper block
1092 least_freq = LCA_freq;
1093 start_latency = start_lat;
1094 end_latency = end_lat;
1095 if (target <= start_lat)
1096 in_latency = true;
1097 }
1098 }
1100 #ifndef PRODUCT
1101 if (trace_opto_pipelining()) {
1102 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1103 least->_pre_order, start_latency, least_freq);
1104 }
1105 #endif
1107 // See if the latency needs to be updated
1108 if (target < end_latency) {
1109 #ifndef PRODUCT
1110 if (trace_opto_pipelining()) {
1111 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1112 }
1113 #endif
1114 _node_latency->at_put_grow(self->_idx, end_latency);
1115 partial_latency_of_defs(self);
1116 }
1118 return least;
1119 }
1122 //------------------------------schedule_late-----------------------------------
1123 // Now schedule all codes as LATE as possible. This is the LCA in the
1124 // dominator tree of all USES of a value. Pick the block with the least
1125 // loop nesting depth that is lowest in the dominator tree.
1126 extern const char must_clone[];
1127 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) {
1128 #ifndef PRODUCT
1129 if (trace_opto_pipelining())
1130 tty->print("\n#---- schedule_late ----\n");
1131 #endif
1133 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
1134 Node *self;
1136 // Walk over all the nodes from last to first
1137 while (self = iter.next()) {
1138 Block* early = _bbs[self->_idx]; // Earliest legal placement
1140 if (self->is_top()) {
1141 // Top node goes in bb #2 with other constants.
1142 // It must be special-cased, because it has no out edges.
1143 early->add_inst(self);
1144 continue;
1145 }
1147 // No uses, just terminate
1148 if (self->outcnt() == 0) {
1149 assert(self->is_MachProj(), "sanity");
1150 continue; // Must be a dead machine projection
1151 }
1153 // If node is pinned in the block, then no scheduling can be done.
1154 if( self->pinned() ) // Pinned in block?
1155 continue;
1157 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1158 if (mach) {
1159 switch (mach->ideal_Opcode()) {
1160 case Op_CreateEx:
1161 // Don't move exception creation
1162 early->add_inst(self);
1163 continue;
1164 break;
1165 case Op_CheckCastPP:
1166 // Don't move CheckCastPP nodes away from their input, if the input
1167 // is a rawptr (5071820).
1168 Node *def = self->in(1);
1169 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1170 early->add_inst(self);
1171 #ifdef ASSERT
1172 _raw_oops.push(def);
1173 #endif
1174 continue;
1175 }
1176 break;
1177 }
1178 }
1180 // Gather LCA of all uses
1181 Block *LCA = NULL;
1182 {
1183 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1184 // For all uses, find LCA
1185 Node* use = self->fast_out(i);
1186 LCA = raise_LCA_above_use(LCA, use, self, _bbs);
1187 }
1188 } // (Hide defs of imax, i from rest of block.)
1190 // Place temps in the block of their use. This isn't a
1191 // requirement for correctness but it reduces useless
1192 // interference between temps and other nodes.
1193 if (mach != NULL && mach->is_MachTemp()) {
1194 _bbs.map(self->_idx, LCA);
1195 LCA->add_inst(self);
1196 continue;
1197 }
1199 // Check if 'self' could be anti-dependent on memory
1200 if (self->needs_anti_dependence_check()) {
1201 // Hoist LCA above possible-defs and insert anti-dependences to
1202 // defs in new LCA block.
1203 LCA = insert_anti_dependences(LCA, self);
1204 }
1206 if (early->_dom_depth > LCA->_dom_depth) {
1207 // Somehow the LCA has moved above the earliest legal point.
1208 // (One way this can happen is via memory_early_block.)
1209 if (C->subsume_loads() == true && !C->failing()) {
1210 // Retry with subsume_loads == false
1211 // If this is the first failure, the sentinel string will "stick"
1212 // to the Compile object, and the C2Compiler will see it and retry.
1213 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1214 } else {
1215 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1216 C->record_method_not_compilable("late schedule failed: incorrect graph");
1217 }
1218 return;
1219 }
1221 // If there is no opportunity to hoist, then we're done.
1222 // In stress mode, try to hoist even the single operations.
1223 bool try_to_hoist = StressGCM || (LCA != early);
1225 // Must clone guys stay next to use; no hoisting allowed.
1226 // Also cannot hoist guys that alter memory or are otherwise not
1227 // allocatable (hoisting can make a value live longer, leading to
1228 // anti and output dependency problems which are normally resolved
1229 // by the register allocator giving everyone a different register).
1230 if (mach != NULL && must_clone[mach->ideal_Opcode()])
1231 try_to_hoist = false;
1233 Block* late = NULL;
1234 if (try_to_hoist) {
1235 // Now find the block with the least execution frequency.
1236 // Start at the latest schedule and work up to the earliest schedule
1237 // in the dominator tree. Thus the Node will dominate all its uses.
1238 late = hoist_to_cheaper_block(LCA, early, self);
1239 } else {
1240 // Just use the LCA of the uses.
1241 late = LCA;
1242 }
1244 // Put the node into target block
1245 schedule_node_into_block(self, late);
1247 #ifdef ASSERT
1248 if (self->needs_anti_dependence_check()) {
1249 // since precedence edges are only inserted when we're sure they
1250 // are needed make sure that after placement in a block we don't
1251 // need any new precedence edges.
1252 verify_anti_dependences(late, self);
1253 }
1254 #endif
1255 } // Loop until all nodes have been visited
1257 } // end ScheduleLate
1259 //------------------------------GlobalCodeMotion-------------------------------
1260 void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_list ) {
1261 ResourceMark rm;
1263 #ifndef PRODUCT
1264 if (trace_opto_pipelining()) {
1265 tty->print("\n---- Start GlobalCodeMotion ----\n");
1266 }
1267 #endif
1269 // Initialize the bbs.map for things on the proj_list
1270 uint i;
1271 for( i=0; i < proj_list.size(); i++ )
1272 _bbs.map(proj_list[i]->_idx, NULL);
1274 // Set the basic block for Nodes pinned into blocks
1275 Arena *a = Thread::current()->resource_area();
1276 VectorSet visited(a);
1277 schedule_pinned_nodes( visited );
1279 // Find the earliest Block any instruction can be placed in. Some
1280 // instructions are pinned into Blocks. Unpinned instructions can
1281 // appear in last block in which all their inputs occur.
1282 visited.Clear();
1283 Node_List stack(a);
1284 stack.map( (unique >> 1) + 16, NULL); // Pre-grow the list
1285 if (!schedule_early(visited, stack)) {
1286 // Bailout without retry
1287 C->record_method_not_compilable("early schedule failed");
1288 return;
1289 }
1291 // Build Def-Use edges.
1292 proj_list.push(_root); // Add real root as another root
1293 proj_list.pop();
1295 // Compute the latency information (via backwards walk) for all the
1296 // instructions in the graph
1297 _node_latency = new GrowableArray<uint>(); // resource_area allocation
1299 if( C->do_scheduling() )
1300 ComputeLatenciesBackwards(visited, stack);
1302 // Now schedule all codes as LATE as possible. This is the LCA in the
1303 // dominator tree of all USES of a value. Pick the block with the least
1304 // loop nesting depth that is lowest in the dominator tree.
1305 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
1306 schedule_late(visited, stack);
1307 if( C->failing() ) {
1308 // schedule_late fails only when graph is incorrect.
1309 assert(!VerifyGraphEdges, "verification should have failed");
1310 return;
1311 }
1313 unique = C->unique();
1315 #ifndef PRODUCT
1316 if (trace_opto_pipelining()) {
1317 tty->print("\n---- Detect implicit null checks ----\n");
1318 }
1319 #endif
1321 // Detect implicit-null-check opportunities. Basically, find NULL checks
1322 // with suitable memory ops nearby. Use the memory op to do the NULL check.
1323 // I can generate a memory op if there is not one nearby.
1324 if (C->is_method_compilation()) {
1325 // Don't do it for natives, adapters, or runtime stubs
1326 int allowed_reasons = 0;
1327 // ...and don't do it when there have been too many traps, globally.
1328 for (int reason = (int)Deoptimization::Reason_none+1;
1329 reason < Compile::trapHistLength; reason++) {
1330 assert(reason < BitsPerInt, "recode bit map");
1331 if (!C->too_many_traps((Deoptimization::DeoptReason) reason))
1332 allowed_reasons |= nth_bit(reason);
1333 }
1334 // By reversing the loop direction we get a very minor gain on mpegaudio.
1335 // Feel free to revert to a forward loop for clarity.
1336 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1337 for( int i= matcher._null_check_tests.size()-2; i>=0; i-=2 ) {
1338 Node *proj = matcher._null_check_tests[i ];
1339 Node *val = matcher._null_check_tests[i+1];
1340 _bbs[proj->_idx]->implicit_null_check(this, proj, val, allowed_reasons);
1341 // The implicit_null_check will only perform the transformation
1342 // if the null branch is truly uncommon, *and* it leads to an
1343 // uncommon trap. Combined with the too_many_traps guards
1344 // above, this prevents SEGV storms reported in 6366351,
1345 // by recompiling offending methods without this optimization.
1346 }
1347 }
1349 #ifndef PRODUCT
1350 if (trace_opto_pipelining()) {
1351 tty->print("\n---- Start Local Scheduling ----\n");
1352 }
1353 #endif
1355 // Schedule locally. Right now a simple topological sort.
1356 // Later, do a real latency aware scheduler.
1357 uint max_idx = C->unique();
1358 GrowableArray<int> ready_cnt(max_idx, max_idx, -1);
1359 visited.Clear();
1360 for (i = 0; i < _num_blocks; i++) {
1361 if (!_blocks[i]->schedule_local(this, matcher, ready_cnt, visited)) {
1362 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1363 C->record_method_not_compilable("local schedule failed");
1364 }
1365 return;
1366 }
1367 }
1369 // If we inserted any instructions between a Call and his CatchNode,
1370 // clone the instructions on all paths below the Catch.
1371 for( i=0; i < _num_blocks; i++ )
1372 _blocks[i]->call_catch_cleanup(_bbs, C);
1374 #ifndef PRODUCT
1375 if (trace_opto_pipelining()) {
1376 tty->print("\n---- After GlobalCodeMotion ----\n");
1377 for (uint i = 0; i < _num_blocks; i++) {
1378 _blocks[i]->dump();
1379 }
1380 }
1381 #endif
1382 // Dead.
1383 _node_latency = (GrowableArray<uint> *)0xdeadbeef;
1384 }
1387 //------------------------------Estimate_Block_Frequency-----------------------
1388 // Estimate block frequencies based on IfNode probabilities.
1389 void PhaseCFG::Estimate_Block_Frequency() {
1391 // Force conditional branches leading to uncommon traps to be unlikely,
1392 // not because we get to the uncommon_trap with less relative frequency,
1393 // but because an uncommon_trap typically causes a deopt, so we only get
1394 // there once.
1395 if (C->do_freq_based_layout()) {
1396 Block_List worklist;
1397 Block* root_blk = _blocks[0];
1398 for (uint i = 1; i < root_blk->num_preds(); i++) {
1399 Block *pb = _bbs[root_blk->pred(i)->_idx];
1400 if (pb->has_uncommon_code()) {
1401 worklist.push(pb);
1402 }
1403 }
1404 while (worklist.size() > 0) {
1405 Block* uct = worklist.pop();
1406 if (uct == _broot) continue;
1407 for (uint i = 1; i < uct->num_preds(); i++) {
1408 Block *pb = _bbs[uct->pred(i)->_idx];
1409 if (pb->_num_succs == 1) {
1410 worklist.push(pb);
1411 } else if (pb->num_fall_throughs() == 2) {
1412 pb->update_uncommon_branch(uct);
1413 }
1414 }
1415 }
1416 }
1418 // Create the loop tree and calculate loop depth.
1419 _root_loop = create_loop_tree();
1420 _root_loop->compute_loop_depth(0);
1422 // Compute block frequency of each block, relative to a single loop entry.
1423 _root_loop->compute_freq();
1425 // Adjust all frequencies to be relative to a single method entry
1426 _root_loop->_freq = 1.0;
1427 _root_loop->scale_freq();
1429 // Save outmost loop frequency for LRG frequency threshold
1430 _outer_loop_freq = _root_loop->outer_loop_freq();
1432 // force paths ending at uncommon traps to be infrequent
1433 if (!C->do_freq_based_layout()) {
1434 Block_List worklist;
1435 Block* root_blk = _blocks[0];
1436 for (uint i = 1; i < root_blk->num_preds(); i++) {
1437 Block *pb = _bbs[root_blk->pred(i)->_idx];
1438 if (pb->has_uncommon_code()) {
1439 worklist.push(pb);
1440 }
1441 }
1442 while (worklist.size() > 0) {
1443 Block* uct = worklist.pop();
1444 uct->_freq = PROB_MIN;
1445 for (uint i = 1; i < uct->num_preds(); i++) {
1446 Block *pb = _bbs[uct->pred(i)->_idx];
1447 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1448 worklist.push(pb);
1449 }
1450 }
1451 }
1452 }
1454 #ifdef ASSERT
1455 for (uint i = 0; i < _num_blocks; i++ ) {
1456 Block *b = _blocks[i];
1457 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1458 }
1459 #endif
1461 #ifndef PRODUCT
1462 if (PrintCFGBlockFreq) {
1463 tty->print_cr("CFG Block Frequencies");
1464 _root_loop->dump_tree();
1465 if (Verbose) {
1466 tty->print_cr("PhaseCFG dump");
1467 dump();
1468 tty->print_cr("Node dump");
1469 _root->dump(99999);
1470 }
1471 }
1472 #endif
1473 }
1475 //----------------------------create_loop_tree--------------------------------
1476 // Create a loop tree from the CFG
1477 CFGLoop* PhaseCFG::create_loop_tree() {
1479 #ifdef ASSERT
1480 assert( _blocks[0] == _broot, "" );
1481 for (uint i = 0; i < _num_blocks; i++ ) {
1482 Block *b = _blocks[i];
1483 // Check that _loop field are clear...we could clear them if not.
1484 assert(b->_loop == NULL, "clear _loop expected");
1485 // Sanity check that the RPO numbering is reflected in the _blocks array.
1486 // It doesn't have to be for the loop tree to be built, but if it is not,
1487 // then the blocks have been reordered since dom graph building...which
1488 // may question the RPO numbering
1489 assert(b->_rpo == i, "unexpected reverse post order number");
1490 }
1491 #endif
1493 int idct = 0;
1494 CFGLoop* root_loop = new CFGLoop(idct++);
1496 Block_List worklist;
1498 // Assign blocks to loops
1499 for(uint i = _num_blocks - 1; i > 0; i-- ) { // skip Root block
1500 Block *b = _blocks[i];
1502 if (b->head()->is_Loop()) {
1503 Block* loop_head = b;
1504 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1505 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1506 Block* tail = _bbs[tail_n->_idx];
1508 // Defensively filter out Loop nodes for non-single-entry loops.
1509 // For all reasonable loops, the head occurs before the tail in RPO.
1510 if (i <= tail->_rpo) {
1512 // The tail and (recursive) predecessors of the tail
1513 // are made members of a new loop.
1515 assert(worklist.size() == 0, "nonempty worklist");
1516 CFGLoop* nloop = new CFGLoop(idct++);
1517 assert(loop_head->_loop == NULL, "just checking");
1518 loop_head->_loop = nloop;
1519 // Add to nloop so push_pred() will skip over inner loops
1520 nloop->add_member(loop_head);
1521 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, _bbs);
1523 while (worklist.size() > 0) {
1524 Block* member = worklist.pop();
1525 if (member != loop_head) {
1526 for (uint j = 1; j < member->num_preds(); j++) {
1527 nloop->push_pred(member, j, worklist, _bbs);
1528 }
1529 }
1530 }
1531 }
1532 }
1533 }
1535 // Create a member list for each loop consisting
1536 // of both blocks and (immediate child) loops.
1537 for (uint i = 0; i < _num_blocks; i++) {
1538 Block *b = _blocks[i];
1539 CFGLoop* lp = b->_loop;
1540 if (lp == NULL) {
1541 // Not assigned to a loop. Add it to the method's pseudo loop.
1542 b->_loop = root_loop;
1543 lp = root_loop;
1544 }
1545 if (lp == root_loop || b != lp->head()) { // loop heads are already members
1546 lp->add_member(b);
1547 }
1548 if (lp != root_loop) {
1549 if (lp->parent() == NULL) {
1550 // Not a nested loop. Make it a child of the method's pseudo loop.
1551 root_loop->add_nested_loop(lp);
1552 }
1553 if (b == lp->head()) {
1554 // Add nested loop to member list of parent loop.
1555 lp->parent()->add_member(lp);
1556 }
1557 }
1558 }
1560 return root_loop;
1561 }
1563 //------------------------------push_pred--------------------------------------
1564 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk) {
1565 Node* pred_n = blk->pred(i);
1566 Block* pred = node_to_blk[pred_n->_idx];
1567 CFGLoop *pred_loop = pred->_loop;
1568 if (pred_loop == NULL) {
1569 // Filter out blocks for non-single-entry loops.
1570 // For all reasonable loops, the head occurs before the tail in RPO.
1571 if (pred->_rpo > head()->_rpo) {
1572 pred->_loop = this;
1573 worklist.push(pred);
1574 }
1575 } else if (pred_loop != this) {
1576 // Nested loop.
1577 while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1578 pred_loop = pred_loop->_parent;
1579 }
1580 // Make pred's loop be a child
1581 if (pred_loop->_parent == NULL) {
1582 add_nested_loop(pred_loop);
1583 // Continue with loop entry predecessor.
1584 Block* pred_head = pred_loop->head();
1585 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1586 assert(pred_head != head(), "loop head in only one loop");
1587 push_pred(pred_head, LoopNode::EntryControl, worklist, node_to_blk);
1588 } else {
1589 assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1590 }
1591 }
1592 }
1594 //------------------------------add_nested_loop--------------------------------
1595 // Make cl a child of the current loop in the loop tree.
1596 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1597 assert(_parent == NULL, "no parent yet");
1598 assert(cl != this, "not my own parent");
1599 cl->_parent = this;
1600 CFGLoop* ch = _child;
1601 if (ch == NULL) {
1602 _child = cl;
1603 } else {
1604 while (ch->_sibling != NULL) { ch = ch->_sibling; }
1605 ch->_sibling = cl;
1606 }
1607 }
1609 //------------------------------compute_loop_depth-----------------------------
1610 // Store the loop depth in each CFGLoop object.
1611 // Recursively walk the children to do the same for them.
1612 void CFGLoop::compute_loop_depth(int depth) {
1613 _depth = depth;
1614 CFGLoop* ch = _child;
1615 while (ch != NULL) {
1616 ch->compute_loop_depth(depth + 1);
1617 ch = ch->_sibling;
1618 }
1619 }
1621 //------------------------------compute_freq-----------------------------------
1622 // Compute the frequency of each block and loop, relative to a single entry
1623 // into the dominating loop head.
1624 void CFGLoop::compute_freq() {
1625 // Bottom up traversal of loop tree (visit inner loops first.)
1626 // Set loop head frequency to 1.0, then transitively
1627 // compute frequency for all successors in the loop,
1628 // as well as for each exit edge. Inner loops are
1629 // treated as single blocks with loop exit targets
1630 // as the successor blocks.
1632 // Nested loops first
1633 CFGLoop* ch = _child;
1634 while (ch != NULL) {
1635 ch->compute_freq();
1636 ch = ch->_sibling;
1637 }
1638 assert (_members.length() > 0, "no empty loops");
1639 Block* hd = head();
1640 hd->_freq = 1.0f;
1641 for (int i = 0; i < _members.length(); i++) {
1642 CFGElement* s = _members.at(i);
1643 float freq = s->_freq;
1644 if (s->is_block()) {
1645 Block* b = s->as_Block();
1646 for (uint j = 0; j < b->_num_succs; j++) {
1647 Block* sb = b->_succs[j];
1648 update_succ_freq(sb, freq * b->succ_prob(j));
1649 }
1650 } else {
1651 CFGLoop* lp = s->as_CFGLoop();
1652 assert(lp->_parent == this, "immediate child");
1653 for (int k = 0; k < lp->_exits.length(); k++) {
1654 Block* eb = lp->_exits.at(k).get_target();
1655 float prob = lp->_exits.at(k).get_prob();
1656 update_succ_freq(eb, freq * prob);
1657 }
1658 }
1659 }
1661 // For all loops other than the outer, "method" loop,
1662 // sum and normalize the exit probability. The "method" loop
1663 // should keep the initial exit probability of 1, so that
1664 // inner blocks do not get erroneously scaled.
1665 if (_depth != 0) {
1666 // Total the exit probabilities for this loop.
1667 float exits_sum = 0.0f;
1668 for (int i = 0; i < _exits.length(); i++) {
1669 exits_sum += _exits.at(i).get_prob();
1670 }
1672 // Normalize the exit probabilities. Until now, the
1673 // probabilities estimate the possibility of exit per
1674 // a single loop iteration; afterward, they estimate
1675 // the probability of exit per loop entry.
1676 for (int i = 0; i < _exits.length(); i++) {
1677 Block* et = _exits.at(i).get_target();
1678 float new_prob = 0.0f;
1679 if (_exits.at(i).get_prob() > 0.0f) {
1680 new_prob = _exits.at(i).get_prob() / exits_sum;
1681 }
1682 BlockProbPair bpp(et, new_prob);
1683 _exits.at_put(i, bpp);
1684 }
1686 // Save the total, but guard against unreasonable probability,
1687 // as the value is used to estimate the loop trip count.
1688 // An infinite trip count would blur relative block
1689 // frequencies.
1690 if (exits_sum > 1.0f) exits_sum = 1.0;
1691 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1692 _exit_prob = exits_sum;
1693 }
1694 }
1696 //------------------------------succ_prob-------------------------------------
1697 // Determine the probability of reaching successor 'i' from the receiver block.
1698 float Block::succ_prob(uint i) {
1699 int eidx = end_idx();
1700 Node *n = _nodes[eidx]; // Get ending Node
1702 int op = n->Opcode();
1703 if (n->is_Mach()) {
1704 if (n->is_MachNullCheck()) {
1705 // Can only reach here if called after lcm. The original Op_If is gone,
1706 // so we attempt to infer the probability from one or both of the
1707 // successor blocks.
1708 assert(_num_succs == 2, "expecting 2 successors of a null check");
1709 // If either successor has only one predecessor, then the
1710 // probability estimate can be derived using the
1711 // relative frequency of the successor and this block.
1712 if (_succs[i]->num_preds() == 2) {
1713 return _succs[i]->_freq / _freq;
1714 } else if (_succs[1-i]->num_preds() == 2) {
1715 return 1 - (_succs[1-i]->_freq / _freq);
1716 } else {
1717 // Estimate using both successor frequencies
1718 float freq = _succs[i]->_freq;
1719 return freq / (freq + _succs[1-i]->_freq);
1720 }
1721 }
1722 op = n->as_Mach()->ideal_Opcode();
1723 }
1726 // Switch on branch type
1727 switch( op ) {
1728 case Op_CountedLoopEnd:
1729 case Op_If: {
1730 assert (i < 2, "just checking");
1731 // Conditionals pass on only part of their frequency
1732 float prob = n->as_MachIf()->_prob;
1733 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1734 // If succ[i] is the FALSE branch, invert path info
1735 if( _nodes[i + eidx + 1]->Opcode() == Op_IfFalse ) {
1736 return 1.0f - prob; // not taken
1737 } else {
1738 return prob; // taken
1739 }
1740 }
1742 case Op_Jump:
1743 // Divide the frequency between all successors evenly
1744 return 1.0f/_num_succs;
1746 case Op_Catch: {
1747 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1748 if (ci->_con == CatchProjNode::fall_through_index) {
1749 // Fall-thru path gets the lion's share.
1750 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1751 } else {
1752 // Presume exceptional paths are equally unlikely
1753 return PROB_UNLIKELY_MAG(5);
1754 }
1755 }
1757 case Op_Root:
1758 case Op_Goto:
1759 // Pass frequency straight thru to target
1760 return 1.0f;
1762 case Op_NeverBranch:
1763 return 0.0f;
1765 case Op_TailCall:
1766 case Op_TailJump:
1767 case Op_Return:
1768 case Op_Halt:
1769 case Op_Rethrow:
1770 // Do not push out freq to root block
1771 return 0.0f;
1773 default:
1774 ShouldNotReachHere();
1775 }
1777 return 0.0f;
1778 }
1780 //------------------------------num_fall_throughs-----------------------------
1781 // Return the number of fall-through candidates for a block
1782 int Block::num_fall_throughs() {
1783 int eidx = end_idx();
1784 Node *n = _nodes[eidx]; // Get ending Node
1786 int op = n->Opcode();
1787 if (n->is_Mach()) {
1788 if (n->is_MachNullCheck()) {
1789 // In theory, either side can fall-thru, for simplicity sake,
1790 // let's say only the false branch can now.
1791 return 1;
1792 }
1793 op = n->as_Mach()->ideal_Opcode();
1794 }
1796 // Switch on branch type
1797 switch( op ) {
1798 case Op_CountedLoopEnd:
1799 case Op_If:
1800 return 2;
1802 case Op_Root:
1803 case Op_Goto:
1804 return 1;
1806 case Op_Catch: {
1807 for (uint i = 0; i < _num_succs; i++) {
1808 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1809 if (ci->_con == CatchProjNode::fall_through_index) {
1810 return 1;
1811 }
1812 }
1813 return 0;
1814 }
1816 case Op_Jump:
1817 case Op_NeverBranch:
1818 case Op_TailCall:
1819 case Op_TailJump:
1820 case Op_Return:
1821 case Op_Halt:
1822 case Op_Rethrow:
1823 return 0;
1825 default:
1826 ShouldNotReachHere();
1827 }
1829 return 0;
1830 }
1832 //------------------------------succ_fall_through-----------------------------
1833 // Return true if a specific successor could be fall-through target.
1834 bool Block::succ_fall_through(uint i) {
1835 int eidx = end_idx();
1836 Node *n = _nodes[eidx]; // Get ending Node
1838 int op = n->Opcode();
1839 if (n->is_Mach()) {
1840 if (n->is_MachNullCheck()) {
1841 // In theory, either side can fall-thru, for simplicity sake,
1842 // let's say only the false branch can now.
1843 return _nodes[i + eidx + 1]->Opcode() == Op_IfFalse;
1844 }
1845 op = n->as_Mach()->ideal_Opcode();
1846 }
1848 // Switch on branch type
1849 switch( op ) {
1850 case Op_CountedLoopEnd:
1851 case Op_If:
1852 case Op_Root:
1853 case Op_Goto:
1854 return true;
1856 case Op_Catch: {
1857 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1858 return ci->_con == CatchProjNode::fall_through_index;
1859 }
1861 case Op_Jump:
1862 case Op_NeverBranch:
1863 case Op_TailCall:
1864 case Op_TailJump:
1865 case Op_Return:
1866 case Op_Halt:
1867 case Op_Rethrow:
1868 return false;
1870 default:
1871 ShouldNotReachHere();
1872 }
1874 return false;
1875 }
1877 //------------------------------update_uncommon_branch------------------------
1878 // Update the probability of a two-branch to be uncommon
1879 void Block::update_uncommon_branch(Block* ub) {
1880 int eidx = end_idx();
1881 Node *n = _nodes[eidx]; // Get ending Node
1883 int op = n->as_Mach()->ideal_Opcode();
1885 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
1886 assert(num_fall_throughs() == 2, "must be a two way branch block");
1888 // Which successor is ub?
1889 uint s;
1890 for (s = 0; s <_num_succs; s++) {
1891 if (_succs[s] == ub) break;
1892 }
1893 assert(s < 2, "uncommon successor must be found");
1895 // If ub is the true path, make the proability small, else
1896 // ub is the false path, and make the probability large
1897 bool invert = (_nodes[s + eidx + 1]->Opcode() == Op_IfFalse);
1899 // Get existing probability
1900 float p = n->as_MachIf()->_prob;
1902 if (invert) p = 1.0 - p;
1903 if (p > PROB_MIN) {
1904 p = PROB_MIN;
1905 }
1906 if (invert) p = 1.0 - p;
1908 n->as_MachIf()->_prob = p;
1909 }
1911 //------------------------------update_succ_freq-------------------------------
1912 // Update the appropriate frequency associated with block 'b', a successor of
1913 // a block in this loop.
1914 void CFGLoop::update_succ_freq(Block* b, float freq) {
1915 if (b->_loop == this) {
1916 if (b == head()) {
1917 // back branch within the loop
1918 // Do nothing now, the loop carried frequency will be
1919 // adjust later in scale_freq().
1920 } else {
1921 // simple branch within the loop
1922 b->_freq += freq;
1923 }
1924 } else if (!in_loop_nest(b)) {
1925 // branch is exit from this loop
1926 BlockProbPair bpp(b, freq);
1927 _exits.append(bpp);
1928 } else {
1929 // branch into nested loop
1930 CFGLoop* ch = b->_loop;
1931 ch->_freq += freq;
1932 }
1933 }
1935 //------------------------------in_loop_nest-----------------------------------
1936 // Determine if block b is in the receiver's loop nest.
1937 bool CFGLoop::in_loop_nest(Block* b) {
1938 int depth = _depth;
1939 CFGLoop* b_loop = b->_loop;
1940 int b_depth = b_loop->_depth;
1941 if (depth == b_depth) {
1942 return true;
1943 }
1944 while (b_depth > depth) {
1945 b_loop = b_loop->_parent;
1946 b_depth = b_loop->_depth;
1947 }
1948 return b_loop == this;
1949 }
1951 //------------------------------scale_freq-------------------------------------
1952 // Scale frequency of loops and blocks by trip counts from outer loops
1953 // Do a top down traversal of loop tree (visit outer loops first.)
1954 void CFGLoop::scale_freq() {
1955 float loop_freq = _freq * trip_count();
1956 _freq = loop_freq;
1957 for (int i = 0; i < _members.length(); i++) {
1958 CFGElement* s = _members.at(i);
1959 float block_freq = s->_freq * loop_freq;
1960 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
1961 block_freq = MIN_BLOCK_FREQUENCY;
1962 s->_freq = block_freq;
1963 }
1964 CFGLoop* ch = _child;
1965 while (ch != NULL) {
1966 ch->scale_freq();
1967 ch = ch->_sibling;
1968 }
1969 }
1971 // Frequency of outer loop
1972 float CFGLoop::outer_loop_freq() const {
1973 if (_child != NULL) {
1974 return _child->_freq;
1975 }
1976 return _freq;
1977 }
1979 #ifndef PRODUCT
1980 //------------------------------dump_tree--------------------------------------
1981 void CFGLoop::dump_tree() const {
1982 dump();
1983 if (_child != NULL) _child->dump_tree();
1984 if (_sibling != NULL) _sibling->dump_tree();
1985 }
1987 //------------------------------dump-------------------------------------------
1988 void CFGLoop::dump() const {
1989 for (int i = 0; i < _depth; i++) tty->print(" ");
1990 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
1991 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
1992 for (int i = 0; i < _depth; i++) tty->print(" ");
1993 tty->print(" members:", _id);
1994 int k = 0;
1995 for (int i = 0; i < _members.length(); i++) {
1996 if (k++ >= 6) {
1997 tty->print("\n ");
1998 for (int j = 0; j < _depth+1; j++) tty->print(" ");
1999 k = 0;
2000 }
2001 CFGElement *s = _members.at(i);
2002 if (s->is_block()) {
2003 Block *b = s->as_Block();
2004 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2005 } else {
2006 CFGLoop* lp = s->as_CFGLoop();
2007 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2008 }
2009 }
2010 tty->print("\n");
2011 for (int i = 0; i < _depth; i++) tty->print(" ");
2012 tty->print(" exits: ");
2013 k = 0;
2014 for (int i = 0; i < _exits.length(); i++) {
2015 if (k++ >= 7) {
2016 tty->print("\n ");
2017 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2018 k = 0;
2019 }
2020 Block *blk = _exits.at(i).get_target();
2021 float prob = _exits.at(i).get_prob();
2022 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2023 }
2024 tty->print("\n");
2025 }
2026 #endif